acceptodds
Under review as a conference paper at ICLR 2027

NeuroMI: Progressive EEG Foundation Modeling for Neural Decoding of Motor Intention

Abstract

Motor imagery (MI) provides a non-invasive paradigm for decoding motor intention-related neural states through brain-computer interfaces (BCIs). However, the practical deployment of MI-based BCIs remains constrained by data scarcity, substantial inter-subject variability, and poor cross-dataset generalization. Existing methods either train task-specific models from scratch on small MI datasets, which compromises robustness, or directly fine-tune generic electroencephalography (EEG) pre-trained models, resulting in a significant domain gap between generic representations and MI-specific neural signatures. This paper proposes a hierarchical three-stage foundation model framework, NeuroMI, to systematically address this gap. In the first stage, a general-purpose EEG encoder is designed and pre-trained on large-scale EEG corpora via self-supervised learning, yielding transferable spatio-temporal-spectral representations. In the second stage, we build a multi-dataset MI-specific EEG corpus (approximately 300,000 samples) and specialize the encoder on it with neurophysiologically motivated objectives, explicitly steering the model toward motor-relevant neural dynamics. In the third stage, the specialized encoder is adapted to downstream MI tasks using lightweight task-specific heads, enabling efficient transfer across subjects and datasets. Experiments on multiple public MI datasets demonstrate that the proposed framework consistently outperforms both generic EEG foundation models and conventional MI-specific architectures. Beyond achieving improved decoding accuracy, the model yields interpretable representations that align with established MI neurophysiology. These results provide a practical pathway toward robust MI decoding in real-world BCI applications.

Primary area:
applications to neuroscience & cognitive science
Keywords:
brain-computer interfaces, motor imagery, EEG, foundation model, motor intention

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_jU4sOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces NeuroMI, a three-stage progressive EEG foundation model for motor imagery (MI) decoding. In Stage 1, the model learns general EEG representations from the TUEG corpus through raw-signal reconstruction and frequency-band energy reconstruction. In Stage 2, it undergoes MI-specific pretraining on approximately 290,000 four-second MI samples, using a classification loss together with auxiliary regression objectives for μ/β-band power, event-related desynchronization (ERD), and filter bank common spatial pattern (FBCSP) features. In Stage 3, the encoder is frozen and only a lightweight classification head is trained. The model combines a multi-scale CNN, frequency-aware positional encoding, and a spatiotemporal fusion Transformer, and is evaluated against task-specific models and EEG foundation models on nine public MI datasets. The results are competitive, but further evidence is needed to substantiate cross-subject generalization and establish the individual contributions of the proposed components.

Critical strengths
  1. The problem is practically relevant. NeuroMI uses progressive pretraining to bridge the gap between general EEG representations and the requirements of MI decoding. Incorporating sensorimotor rhythms, ERD, and FBCSP features as auxiliary supervision offers a useful direction for integrating neurophysiological priors into the domain adaptation of EEG foundation models.
  2. The experimental evaluation is relatively broad. The authors compare the method with several task-specific models and EEG foundation models on nine public MI datasets, and also report ablations of auxiliary losses, feature visualizations, and analyses of training-data scale. These experiments provide some supporting evidence for the proposed approach.
Critical weaknesses
  1. The individual contributions of the pretraining stages and the components within the EEG encoder remain unclear. The observed gains may come from several factors, including the additional MI training data, Stage 1 general-purpose pretraining, Stage 2 classification supervision, neurophysiological auxiliary objectives, and the encoder architecture. The current ablations do not sufficiently disentangle these factors. In particular, the independent contribution of Stage 1 and the roles of architectural components such as frequency-aware positional encoding have not been specifically validated.
  2. The neurophysiological supervision targets and their implementation details require clarification. The relationship between the 1–40 Hz band-pass filter and the 30–50 Hz band-energy target is unclear. The physiological validity of using the first 0.5 seconds of each trial as the ERD baseline should be explained in light of the cue timing in each dataset. For FBCSP, the filter-bank frequency bands, number of CSP components, definition of the 280-dimensional feature vector, and scope of data used to estimate the spatial filters should be specified.
  3. Evidence for generalization is insufficient. The downstream experiments primarily use within-subject five-fold cross-validation, and the classification head is still trained on labeled samples from the target dataset. This setup neither adequately evaluates generalization to unseen subjects nor establishes zero-shot cross-dataset transfer without labeled target-domain data.
Questions for authors
  1. Cross-subject generalization and cross-dataset transfer. Add a leave-one-subject-out or another subject-independent evaluation, as well as a zero-shot cross-dataset transfer experiment in which no labeled samples from the target dataset are used to train the classification head should be better. If these evaluations are outside the scope of the paper, how should the generalization claims supported by the current results be delimited?
  2. Individual contributions of progressive pretraining and the architecture. In order to determine the independent contributions of the two stages, it is necessary to compare the performance of completely removing either stage one or stage two. In addition, providing ablations of key architectural components is also important, such as frequency-aware positional encoding, and clarifying differences among the foundation-model baselines in their pretraining data, supervision signals, and training configurations.
  3. Implementation of the neurophysiological auxiliary objectives. How is the Stage 1 band-energy target covering 30–50 Hz related to the 1–40 Hz filtering range? Does the first 0.5 seconds of each trial correspond to a pre-task baseline for ERD in every dataset? For FBCSP, please specify the filter-bank frequency bands, the number of CSP components, the definition of the 280-dimensional feature vector, and how the spatial filters are estimated, including which data are used for estimation.
  4. Model adaptability. As a EEG Foundation Model, can the model truly accommodate different EEG channel configurations and input durations?
  5. It seems that MIRepNet is more competitive on most datasets. Why isn't it included in Figure 1(b)?
Reviewer_jjbNOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This work pretrains a CNN-patch Transformer encoder on TUEG, specializes it on a curated multi-dataset MI corpus of about 290k trials with a classification loss plus regression of mu/beta band power, ERD, and FBCSP features, and then freezes it and trains a small head per subject. On nine public MI datasets under within-subject 5-fold CV, NeuroMI is reported best on every dataset against task-specific models and EEG foundation models, with ablation studies of the Stage 2 losses and interpretability figures.

Critical strengths
  1. The curated MI corpus (about 290k trials from public MI datasets with unified preprocessing) is a useful resource in itself.
  2. Specializing a general EEG encoder toward MI before freezing it is a reasonable middle ground between generic foundation models and per-subject models, and the training details are given.
Critical weaknesses
  1. The evaluation protocol does not match the claims. This work is motivated by inter-subject variability, but all results are within-subject stratified 5-fold CV (Table D.3). Trials from the same session are mixed across folds, so this is neither cross-subject nor time-respecting. LOSO or at least cross-session results are needed. The "linear probe" is also a two-layer MLP trained per subject, so subject-specific tuning is still required.
  2. The result is closer to a task-specific model than a foundation model. Regressing mu/beta, ERD, and FBCSP targets builds hand-crafted MI priors into the encoder, and nothing shows the encoder learned more than those priors. The backbone is a standard Transformer on CNN patches although state-space EEG encoders exist and are cited (EEGMamba), and the ablation studies cover only the losses, not any architectural component. Eq. 1 uses undefined symbols, Fig. 2 is unreadable, and Figs. 3-4 do not say which data or layer they use.
  3. The comparison is confounded. NeuroMI sees about 300k labeled MI trials in Stage 2, while the foundation-model baselines did not see this corpus and the task-specific baselines train on one subject. Table 2 suggests most of the gain comes from the classification loss on this corpus, with the mechanism losses adding little within the standard deviations. Several foundation-model baselines are near chance on binary tasks, which suggests the unified 15-channel, 200 Hz template does not fit their expected inputs. A baseline fine-tuned on the same corpus with the classification loss only is the fair control.
Questions for authors
  1. How does NeuroMI perform under LOSO or cross-session evaluation, compared with the strongest baselines?
  2. Were sessions kept separate when forming the folds within a subject?
  3. How much of the gain remains over a baseline (e.g., EEG Conformer or CBraMod) fine-tuned on the same corpus with the classification loss only?
  4. Which dataset, subjects, and Grad-CAM layer are used in Figs. 3 and 4?
Reviewer_9RrFOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

NeuroMI is a further contribution to the rapidly growing literature on so-called EEG foundation models. NeuroMI focusses on motor imagery (MI) decoding. The proposed two-step approach combines general EEG pretraining with subsequent supervised specialization on a collection of MI-related datasets, followed by task-specific training. The architecture / approach combines multiscale convolutional embeddings, spectral information, and spatial and temporal attention. During specialization, classification is supplemented by prediction of mu/beta band power, the so-called event-related desynchronization (ERD), and filter-bank common spatial pattern (FBCSP) features. Evaluation on nine downstream datasets reports higher performance than the tested task-specific and foundation-model baselines. Additional analyses investigate the auxiliary objectives and the physiological properties of the learned representations. The work addresses a relevant practical EEG-decoding problem and reports good-looking results.

Critical strengths

None

Critical weaknesses

A major weakness I would see critical is the relatively incremental nature of the NeuroMi contribution. Multiscale convolutions, spectral features, spatial and temporal attention, masked reconstruction, and auxiliary feature prediction - alll of them are established design ideas. Related EEG "foundation model" architectures already combine spatial/temporal attention with temporal and spectral information.

More specifically, STEAM already proposes a spatio-temporal alignment mixture-of-experts model with hierarchical pre-training for EEG decoding, including MI, and downstream adaptation (Chen et al., arxiv.org/abs/2608.02070). This closely related work is not discussed. NeuroMI differs in some aspects to STEAM, but the core contribution of NeuroMI (the progressive training principle itself) does, on this omitted related work background, not appear to constitute a new contribution.

For such a contribution as NeuroMi, it would be especially important to at least firmly establish the incremental value of the proposed components. However, the central progressive training claim is not fully supported by the necessary downstream / ablation controls. The reported experiments leave open how much of the reported gain actually requires the progressive training procedure and architectural additions, and how much can be explained by supervised training on a large collection of MI data.

Crucial controls are missing:

  • Skip general pretraining: does Stage 1 actually help?
  • Skip MI specialization: does Stage 2 actually help?

The evidence for performance superiority is also critically limited. Five-fold cross-validation and mean +/- standard deviation are reported, but I could not find an evaluation over multiple independent random seeds, either in the manuscript or supplement. It is not lcear whether the reported performance gains are just lucky seeds, or truely reflect statistically significant differences.

In summary, therefore, I see a potentially useful combination of existing methods, but an rather incremental methodological contribution with insufficient evidence for central claims, including the claimed preformance gains. In its present form, I do not consider the contribution strong enough for acceptance at at a highest-level venue such as ICLR, especially also due to the existing but here ignored prior work in STEAM, which in contrast to NeuroMI offers stronger evidence for the usefulness of progressive training by testing the individual stages directly and over multiple runs.

Questions for authors

What was the optimization approach / budget for NeuroMI - was it optimized on the same task which are reported in the manuscript, and if yes, to which extend? In particular, what were the model selection and hyperparameter optimization budgets for NeuroMI and the baselines? Were architecture and auxiliary-loss choices developed using the nine reported downstream benchmarks? If so, which data were used for these choices, and how was independence of the final evaluation ensured?

How many independent training runs were performed for NeuroMI, the baselines, and the ablations? What exactly is the population over which the reported standard deviations were calculated?

Could the authors provide the crucial controls:

  • Skip general pretraining: does Stage 1 actually help?
  • Skip MI specialization: does Stage 2 actually help?

What is the uncertainty across multiple randomly selected random seeds - are the reported performance gains real?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.